tencentdb agent memory getting open-sourced made me rethink agent memory a bit.​ what i like most is its short-term context cleanup. agent runs get messy fast: tool logs, retries, failed branches, repeated observations, and a lot of stuff you probably don’t want dumped back into the prompt.​ tencentdb’s mermaid-style canvas feels practical here. it compresses a messy run into something easier to inspect, while node_id still lets you trace back to the raw data. the claimed token saving, up to 61.38%, is also meaningful if you are running agents on real tasks.​ i also like that it is not just one giant vector db. conversation records, atomic facts, scenario memory, and profile memory are separated, with sqlite / sqlite-vec and markdown files keeping things fairly local and inspectable.​ so yeah, tencentdb looks strong for short-term memory management.​

but compression is not the same thing as learning.​

if an agent spends an hour debugging docker permissions and finally finds a uid/gid mismatch, i don’t just want a cleaner summary of that run. i want the agent to check uid/gid earlier next time and stop starting with chmod 777.​

that is not just shorter memory. that is a reusable debugging habit.​

this is where memos local plugin 2.0 feels like it is solving a different layer of the problem. its focus seems less about reducing token cost but more about turning execution history into better future behavior. that’s a different view.​ the trace layer keeps the step-level record. the policy layer distills patterns across tasks. the world model stores environment-level knowledge. then useful repeated patterns can become reusable skills.​ that feels closer to long-term agent learning than long-term storage.​

the feedback loop is the part i care about most. if a task fails, i don’t want the system to neatly save that failure and accidentally retrieve the same bad path next week. i want the failed path to become less likely. step-level feedback, task-level feedback, llm scoring, and reward propagation all sound like attempts to make memory actually change future decisions.​

the observability side matters too. tencentdb’s markdown-inspectable memory is nice, but the local plugin having a vite viewer ui, live event stream, and structured logs feels more useful when you are trying to understand why an agent picked a certain policy or skill.​ so i don’t really see tencentdb and memos local plugin as direct competitors.​ tencentdb seems very strong at making memory manageable: compress the messy run, reduce token cost, keep it inspectable, and preserve traceability through node_id in a short-term way. but the local plugin feels more like the long-term answer. it is less about storing or compressing what happened, and more about turning traces, feedback, and repeated patterns into better future behavior.​

to me, tencentdb answers: “how do we manage what just happened?”​​ memos answers: “how do we make the agent stop making the same mistake again?”

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；有直接可用的方法、工具或操作价值；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.63